The Verification Gap: when AI reviews its own code — and the language it writes fastest
OpenAI just let an AI model block its own engineers’ pull requests, no human required. Sonar’s 2026 survey says 96% of developers don’t trust AI code anyway. And a 13-language benchmark of AI coding agents just handed Ruby a win over Rust. Two data points, one week, one question: when the AI is doing the typing, what actually holds up?
01The AI gatekeeper is real now
On The Pragmatic Engineer podcast this week, OpenAI Codex lead Thibault Sottiaux confirmed every PR from an OpenAI engineer now passes a mandatory, automated AI security review — and the model can stop a merge outright, no human override required to enforce it. Meanwhile Sonar’s 2026 State of Code Developer Survey (1,100+ devs) found trust hasn’t caught up to adoption. Run the three scenarios below the way the gatekeeper would.
